ºÚÁÏÍø

Prompt engineering is dead, because LLMs are better than humans in writing prompts. So, rather than learning prompt engineering or bookmarking prompts that might not work anymore with new models, you can ask the AI to write the prompt for you and make sure you also instruct it to ask you clarifying questions before generating it.

Example 1: Here an example that has some confidential information about an ongoing research project, this could be used on Aalto AI Assistant ai.aalto.fi.

I am trying to choose which analysis method I should use with my data (transcribed interviews with company owners in Finland related to their attitude towards AI de-skilling their employees). Please make a prompt for me to be used with [model]. Before generating the prompt, ask me a few questions to refine it.

Example 2: Aalto AI assistant does not have harnesses to connect to the internet, so it cannot perform searches or check the literature. Here a prompt that uses the "deep research" feature in Google Gemini (OpenAI's ChatGPT and Anthropic's Claud have something similar) so that it can retrieve information from the internet. This prompt does not contain any confidential details:

I am a brain imaging researchers, but I am not familiar with genomics and brain imaging literature, please write a prompt to be used with Deep Research on Google Gemini so that it explores relevant literature from the past 5 years to identify important peer reviewed articles that combine genetics and brain imaging. Review articles should also be included. Before generating the prompt, ask me a few questions to refine it.

Caveat: while "deep research" seems amazing the first time you use it, many Aalto researchers who came to our past workshops have reported how it actually tends to hallucinate and report details that are not contained in the cited references. Consider deep research as a tool to discover some of the literature, but then please read the actual papers and performs usual searches with article databases.

Generative AI is very good at translating from one language to another, and one of the most useful cases is the translation from natural English to Python (or whatever programming language you work with). In the CodeRefinery lesson on "" are three ways of working with AI generated code:

1. Full control — manually copy-paste code function by function, review each piece, run it yourself. Lowest risk.
2. IDE with AI extension — AI suggests code inline; you approve line by line or block by block; you run the code. Moderate risk.
3. Full agent mode — you act as project manager; the agent writes and runs code autonomously. Highest risk.

Start with scenario 1. Move to scenario 2 only when you are confident you can review what is generated. For scenario 3 see the next section.

AI agents (especially coding agents) carry a higher risks because they can run independently. At least two types of risks should be mentioned.

• Trust drift: the more you use an agent, the more you accept its suggestions without reviewing them. A practical solution is to force review checkpoints: sessions where you manually audit what the agent has done before proceeding.
• Prompt injection: a malicious instruction embedded in an external file (a README, a dataset description) can hijack the agent's actions.

Before running agents on shared infrastructure, read the and review the risk table there. While most risks listed for HPC systems apply equally on a personal computer.

We also have some instructions on how to use coding agents on your computer. See the for examples and more in-depth coverage.

As you can guess, the general recommendation is to only use Aalto approved AI tools and Recommendation 3 of the guidelines applies here. However, often researchers need to compare various tools or need to use new features that just do not exist in the Aalto ecosystem. If you need to test a non approved tool, please follow these general recommendations:

• Using only fully public data in your interactions (ideally public data that has no personal data)
• Assuming the entire interaction is visible to anyone
• Not disclosing unpublished hypotheses, methods, or findings

In some cases the task you need to do is specific, and does not require a large language model. For example transcription of .

Translations from one language to another can also be run locally. Our IT services and Research Software Engineers are working on those tools, in the meantime you can use Aalto AI Assistant and ask to translate the text for you.

AI can be used for generating images, but since the researcher generally does not hold copyrights over purely AI-generated images, they are not accepted for many scientific purposes, such as scientific publications. Nevertheless, this kind of tool can be helpful for aiding researchers in visual creation. Some acceptable strategies for integrating generative AI into scientific workflows include:

1. Conceptual mapping and aesthetics exploration

   When some concept is too abstract, or when having some trouble in translating complex concepts into tangible representations, instead of starting with a blank whiteboard, researchers can use AI to generate the visual landscape of a concept. You can generate visual metaphors for data, experiment with distinct color palettes, textures, and rendering styles to establish a cohesive identity for a research project or grant proposal.

2. Create multiple variations of a concept

   While the previous point is about exploring abstract ideas, this is about deciding how to visually execute a concept you have already defined. AI is capable of generating the exact same concept across many different art styles. You can tune it to sketch an image close to your target aesthetic, allowing you to rapidly test out different compositions, angles, and framings. Once you find the right visual approach, this AI-generated image can serve as a structural base or reference sketch to manually trace and create your own original vector artwork (e.g., using Inkscape or Illustrator).

3. Fast image generator for presentations

   Sometimes you know exactly the image you want, but you do not have enough time to generate it. Visuals can often make the difference in capturing your audience, whether in a group meeting or a seminar presentation. Because these types of outputs are typically for internal purposes or temporary presentations, you do not necessarily need to hold the copyright. AI is great for generating background graphics, conceptual icons, or decorative elements that improve presentation slides.

4. Scripting for 3D modelling and rendering automation

   Besides image models, AI is best for scripting. You can ask an LLM to write Python scripts for Blender (and similar) to programmatically generate precise geometries, arrange camera angles, or animate molecular structures.

5. Text-to-diagram syntax generation

   Maintaining documentation, experimental setups, and project timelines requires clean and modifiable schematics. LLMs are very proficient at converting text descriptions into diagram code (e.g., Mermaid.js, Graphviz, PlantUML). This approach allows you to create flowcharts, diagrams, and roadmaps using code that can be edited, version-controlled via Git, and rendered natively inside markdown documents or lab notebooks.

Qualitative research has traditionally taken place primarily among the researcher, the data, and colleagues on the research team. What is fundamental to this kind of research is the researcher's interpretations and the responsibility to step back and reflect on their influence on the interpretation of the data. In the methodological literature, this is known as reflexivity, and it has long been considered central to qualitative research. In many qualitative traditions, analysis does not seek a single absolute truth; rather, it involves subjective interpretation of the material, treating it as one form of reality. This subjectivity is an integral part of the analytical process and serves as a resource for the researcher (Braun & Clarke, 2019). What counts as a defensible reading depends on the epistemological stance the researcher takes and the rationale behind the interpretive choices they make.

An AI system based on Large Language Models (LLMs) introduces an additional participant in the analytic exchange, supplying words and generating most of the text produced through a chat interface. However, AI does not function straightforwardly as a fully independent third participant, since it is not a coherent or stable entity, and the models operate as a superposition of multiple possible personas. Each prompt generates responses shaped by multiple layers, including pretraining on extensive text data, fine-tuning with example dialogues, and adjustments based on human feedback (such as RLHF), all of which contribute to the formation of the typical assistant persona. Moreover, the output is also affected by the prompt, session context, and memory. Since researchers interact with these systems in everyday language via chat, they can engage in back-and-forth exchanges that influence their thinking and decisions over time. A qualitative interpretation becomes more defensible when the researcher can account for its shaping, but as meaning becomes co-constructed through prolonged dialogue, maintaining reflexivity can be challenging. None of this makes AI unusable in qualitative work; when used reflexively, it can help you think through your own ideas, reorganize material, or test a reading against alternatives, as long as you consciously retain agency over the interpretive process. Here are some useful tips to keep in mind when doing qualitative analysis with an AI system: 

1. Do your own interpretation first. Read the data you collect and form your own analytical reading before you involve AI, so the interpretation begins as yours.
2. Mind what you put into the system. Aalto’s guidance (/en/services/responsible-use-of-artificial-intelligence-in-the-research-process) is explicit that research data containing personal data should not be entered into external AI systems, so prefer Aalto-approved systems and treat the system’s output as potentially sensitive too. 
3. Check the AI output against your data, and keep your own analysis separate. Test its suggestions against your material rather than treating them as a neutral reading. Because the system produces most of the words, you can easily slide from analyzing your data into curating its outputs, which may make the result harder to defend as your own interpretation.
4. Set the context the model works from. You cannot change the training, but you can set the instructions, the chat history, and what the system remembers. Use them on purpose, for example, to make the model push back on your reading of the data. These settings shape the output but do not fully control it, so use them to help you think and articulate ideas, but not to make interpretive decisions for you.
5. Stay reflexive in both directions by keeping a reflexive memo. Record your own shifting assumptions and position alongside the AI’s role in the analysis. Note the prompts you gave, what it suggested, and what you took up or set aside and why, so that every interpretive choice can be traced to your own reasoning.
6. Keep the chat logs as a record of the exchange itself. Your analytical output, in whatever form it takes, such as text, an artifact, or a document, was produced through the exchange itself, whether the system only replied in chat or acted across tools, files, and other agents you could not fully see. Reread conversations as the analysis progresses, paying attention to how the dialogue shaped your thinking.

Recommended readings

Braun, V., & Clarke, V. (2019). Reflecting on reflexive thematic analysis. Qualitative Research in Sport, Exercise and Health, 11(4), 589–597.

Finlay, L. (2002). Negotiating the swamp: The opportunity and challenge of reflexivity in research practice. Qualitative Research, 2(2), 209–230.

Gulay, E., Picco, E., Glerean, E., & Coupette, C. (2026). Relational dissonance in human-AI interactions: The case of knowledge work. In Proceedings of the 2026 CHI Conference on Human Factors in Computing Systems (CHI ’26). ACM.

Madill, A., Jordan, A., & Shirley, C. (2000). Objectivity and reliability in qualitative analysis: Realist, contextualist and radical constructionist epistemologies. British Journal of Psychology, 91(1), 1–20.


 

A1: Manual checking is basically the only reliable method. A workflow from :

1. If a DOI or PubMed ID is present, verify it resolves to the cited article
2. If there is a mismatch, flag the reference
3. Rule out false positives (typos, abbreviation variants)
4. Search the title across four databases: PubMed, Crossref, OpenAlex, Google Scholar

AI detection tools (which are also based on AI...) are not reliable for this purpose.

A2: Horizon Europe has . ICML ran a with instructive findings on policy violations.

A3: Yes, the researchers who will be most resilient are those who can still reason carefully when AI is unavailable or wrong and are always able to go back to the original sources (fully read the papers, know how to check the documentation of software libraries). As every 7-year-old would tell you: Learning the multiplication tables is still worth doing, even if calculators exist.

A4: Findings in the literature are conflicting. In coding, the domain where AI performs best, productivity gains are measurable, but the cost is a higher volume of unmaintainable code and rapidly rising tool costs. In research more broadly, AI may remove some bottlenecks for individual researchers, but shift the bottleneck to peer review which remains a hard limit on the rate at which research can be validated and published. Productivity gains that bypass review quality are not gains for science.

A5: Some studies report clear impacts on the environment (), on the other hand data centers report that they use green energy and aim at zero-waist by 2030 (see ). In general, the energy costs for large AI systems are higher than running a simple web search on a search engine (that does not use AI...), so if the question you have is something that can be solved with a web search, you should consider doing that rather than asking some AI chatbot. Examples of good search engines that allow you to switch off AI results and are also privacy friendly are: ,  

A6: Academic freedom is one of the core principles of research. You should never feel forced to use AI in doing your work, especially when you know that it causes more harm than good. It is however important to learn how these tools were built and how they can (or not) work for your research case. Using generative AI systems built on data scraped , or might not align with your ethical principles. Consider using AI tools which were built responsibly. Unfortunately this is easier said than done: let’s work on this together!

Research misconduct in Finland (and broadly in academia) means fabrication, falsification, and plagiarism. If not used carefully, generative AI can be the ideal misconduct machine as it can very easily engage with any of the three malpractices.

Fabrication: Generative AI systems fabricate all the time, and those fabrications are often called "hallucinations": they can be plausible-looking text, references, statistics, or data that do not exist. A 2026 study () found that over 2800 peer-reviewed papers published in the past two years contain fabricated citations. ArXiv preprint service now imposes a for manuscripts containing AI-fabricated references or other irrefutable AI slop, and publishers are adopting a similar policy . Using AI to help format or suggest citations without manually verifying each one is a significant risk. This applies even to supposedly "safe" reformatting tasks: an AI that rearranges a reference list can change author order, publication year, or invent a new title or journal name (this is more likely to happen if the list is long). Generative AI is also quite bad at summarising articles: it exhagerates or excessively paraphrases the findings, up to the level of fabrication ().

Falsification: AI tools do not reason in the same way humans do, they just predict the statistically likely next token given their training data. Using AI to produce research findings or interpret results creates a high risk of false conclusions. Quantitative and qualitative analysis must be performed with methods that the researcher can validate and re-run reliably (reproducibility). When AI is used to suggest an analysis approach, always independently re-run or verify the key steps with a known method before trusting the result.

Plagiarism: Large language models are trained on basically all available human text (and sounds and images). They synthesise text without citing sources. A recent study () showed it was possible to extract near-verbatim up to 95% of a Harry Potter book from a large language model. Even if you are not copy-pasting directly from AI generated output to avoid plagiarism, using AI to brainstorm an idea you later present as original, without checking whether that idea already appears in the literature, can constitute plagiarism.